Foundation model for efficient biological discovery in single-molecule time traces.

Area of Science:

• Biophysics
• Molecular Biology
• Computational Biology

Background:

• Single-molecule fluorescence microscopy (SMFM) provides critical biological insights but analyzing time traces for rare events is challenging.
• Manual inspection and ad hoc methods limit the efficiency and objectivity of SMFM data analysis.

Purpose of the Study:

• To develop a systematic and efficient method for discovering rare biological intermediates from SMFM time traces.
• To introduce META-SiM, a foundation model designed to automate and improve SMFM data analysis.

Main Methods:

• Developed META-SiM, a transformer-based foundation model pretrained on diverse SMFM analysis tasks.
• Utilized the META-SiM Projector for visualization and analysis of trace embeddings.
• Combined trace embeddings with local Shannon entropy for identifying subtle, condition-specific behaviors.

Main Results:

• META-SiM demonstrates performance rivaling best-in-class algorithms across various SMFM analysis tasks (classification, segmentation, idealization, photobleaching analysis).
• The META-SiM Projector enables efficient dataset visualization, labeling, comparison, and sharing.
• Application to a single-molecule Förster resonance energy transfer dataset revealed a previously undetected intermediate state in pre-mRNA splicing.

Conclusions:

• META-SiM streamlines SMFM data analysis, removing bottlenecks and enhancing objectivity.
• The model systematizes and accelerates biological discovery from single-molecule data.
• META-SiM facilitates the identification of rare and subtle biological behaviors.